CN115147192A - Recommendation method and recommendation system based on double-view-angle deviation correction - Google Patents

Abstract

The invention discloses a recommendation method based on double-view angle correction, which comprises the following steps: step 1, quantitatively analyzing whether article popularity deviation and user activity deviation exist in data, and constructing a causal graph according with a data generation mechanism; step 2, based on the causal graph constructed in the step 1, a do operator is used in a training stage, P (Y | do (U, I)) is used for replacing traditional P (Y | U, I), the influence of the popularity deviation of the item characteristics and the user characteristics is eliminated, the real preference of the user to the item is expressed, and then a back door adjustment technology is used for estimating a P (Y | do (U, I)) causal estimator through data; step 3, parameterizing the causal estimator in the step 2, so that the causal estimator can be recovered by using existing data and trained, and improving a Bayesian personalized sorting BPR loss function to enable a recommendation model to be capable of carrying out depolarization by using information of two aspects of article popularity and user activity; and 4, in an inference stage, finding out the potentially popular goods on the test set by adjusting the popularity deviation.

Description

Recommendation method and recommendation system based on double-view-angle deviation correction

Technical Field

The invention relates to the field of recommendation systems, in particular to a recommendation method and a recommendation system based on double-view-angle deviation correction.

Background

The advent of recommendation systems is now an era of information explosion, effectively alleviating the problem of information overload. Currently, a recommendation system plays an important role in the fields of e-commerce, social platforms, personalized content recommendation and the like, and becomes an indispensable part in modern internet application scenes.

Conventional recommendation systems have generally focused on finding better models to fit the observed data. However, there are various deviations in the observed data, such as selection deviation, exposure deviation, popularity deviation, etc. [1]. If the observation data is directly used as a signal, the learned model parameters cannot accurately represent the real preference of the user to the article, and the recommendation effect is reduced.

A common recommended bias handling idea is to remove the bias in the training. For example, a common selection bias in the display feedback data, which is due to the user's freedom to select items for scoring, results in an observed score that is not a representative sample of all scores [1]. Existing studies remove the selection bias by correcting the exposure probability of each set of samples using inverse probability weighting. But does not in itself degrade the predictive performance of the model for popularity bias. Instead, in the next recommendation, the exposure probability of popular items is further increased, and the exposure probability of unpopular items is further decreased, resulting in bias amplification that degrades the prediction performance of the model. Ultimately, the personalized experience of the user and the potential revenue of the item provider are also affected.

There is currently a small amount of research to improve the performance of models and user satisfaction by exploiting popularity bias [2]. However, the existing recommendation methods are only considered from the single perspective of the article, and the influence of the user activity on the final click probability under the user perspective is ignored, so that the recommendation result cannot accurately represent the real preference of the user on the article.

In summary, the problem that the recommendation result caused by deviation cannot reflect the real preference of the user commonly exists in the current recommendation system.

Disclosure of Invention

Due to the existence of various deviations in the recommendation system, the recommendation result cannot accurately represent the real preference of the user. In order to solve the problem, the invention provides a recommendation method based on article and user double-view angle correction, which is characterized in that a cause-and-effect graph conforming to a data generation mechanism is constructed by using article and user double-view angles, a cause-and-effect estimator capable of reflecting the real preference of a user is defined based on the cause-and-effect graph in a training stage and is estimated by using data, and a potentially popular article on a test set is found by adjusting popularity in an inference stage, so that the recommendation performance is improved, and therefore, the real preference of the user to the article can be represented more accurately.

The invention provides a recommendation method based on double-view angle deviation correction, wherein the double-view angle refers to a method for simultaneously considering two view angles of an article and a user, and comprises the following steps of:

step 1: quantitatively analyzing whether the data have article popularity deviation and user activity deviation, and constructing a causal graph conforming to a data generation mechanism, for example, as shown in fig. 1;

step 2: based on the causal graph constructed in the step 1, a do operator is used in a training stage, and P (Y | do (U, I)) is adopted to replace traditional P (Y | U, I), so that the influence of popularity deviation of the item characteristics and the user characteristics is eliminated, and the real preference of the user on the item is represented. Estimating a P (Y | do (U, I)) causal estimator through data by using a back door adjustment technology, wherein the causal estimator refers to the real click probability of a user on an article;

and step 3: parameterizing the cause and effect estimator in the step 2, wherein the parameterization means that the cause and effect estimator is represented by a function with parameters, so that the cause and effect estimator can be recovered by using existing data and trained, and a Bayesian Personalized Ranking (BPR) loss function is improved, so that the existing recommendation models, such as a matrix decomposition model and a LightGCN model, can be subjected to depolarization by using information of two aspects, namely commodity popularity and user activity. Specifically, the connecting edges leading to the node I and the node U in FIG. 1 are cut off by P (Y | do (U, I)), thereby cutting off the influence of the confounding factors M and N on the exposure distribution of the article and the user;

and 4, step 4: in the inference stage, potentially popular items on the test set are found by adjusting the popularity bias.

In order to optimize the technical scheme, the specific measures adopted further comprise:

in the step 1, the causal graph comprises a main body composed of a user representation node U, an article representation node I and a final click probability Y, wherein U and I are used as model input and are trained to obtain Y, so that the nodes U and I point to the node Y at the same time. On the basis, nodes M and N are defined and respectively represent article popularity and user activity, M serves as a confounding factor and simultaneously points to nodes I and Y, and N is a confounding factor and simultaneously points to nodes U and Y in the same way. The confounding factors can cause the result Y to have article popularity deviation and user activity deviation, and the invention aims to remove the confounding factors in the training stage and adjust the influence of the confounding factors on the result in the deduction stage.

The article popularity M refers to the article popularity degree, and the popularity of an article i at the time t is defined as

Wherein

Representing the interaction quantity of an item I at the time t, and representing the item set by an item representation node I;

the user activity N refers to the user activity degree, and the activity degree of the user u at the moment t is defined as

Wherein

Representing the interaction quantity of the user U at the time t, wherein the user representation node U represents a set of users;

the user characterization node U represents a user embedded vector;

the item characterization node I represents an item embedding vector;

the final click probability Y represents the click probability of the user on the article after the action of the user and the article.

The method and the device perform quantitative analysis on whether the article popularity deviation exists in the data or not, divide the data set according to the time sequence, divide the data set into ten stages according to the interaction time, and measure the article popularity in different time periods, wherein each stage has the same time interval. On the basis, the popularity offset provided by Yang et al [2] is utilized to depict the variation trend of the popularity of the article along with the time, thereby quantitatively analyzing whether the popularity deviation of the article exists in the data; the popularity offset ranges from [0, ln 2], an item popularity bias is deemed to exist if the value of the popularity offset is greater than zero, and an item popularity bias is deemed to not exist if the popularity offset is equal to zero or very close to zero;

the invention carries out quantitative analysis on whether the user activity deviation exists in the data or not, divides the data set according to the time sequence, divides the data set into ten stages according to the interaction time, and each stage has the same time interval, thereby measuring the user activity in different time periods. Similar to the defined popularity offset, the activity offset is used for describing the change trend of the user activity along with time, and whether the user activity deviation exists in the data can be quantitatively analyzed. The range of the activity offset is [0, ln 2], if the value of the activity offset is greater than zero, the user activity deviation is considered to exist, and if the activity offset is equal to zero or very close to zero, the user activity deviation is considered to not exist.

The item popularity offset is measured by Jensen-Shannon divergence and can be expressed as

Wherein t is ₁ And t ₂ The time periods are two time periods, and the time periods are,

is t ₁ The probability distribution of the item at the moment in time,

is t ₂ Probability distribution of the item at the moment.

The user activityThe offset is also measured using Jensen-Shannon divergence, which can be expressed specifically as

Wherein t is ₁ And t ₂ The two time periods are divided into two time periods,

is t ₁ The probability distribution of the users at the moment in time,

is t ₂ Probability distribution of users at the moment.

In the step 2, the do operator is an intervention behavior for the article representation and the user representation, and aims to remove the popularity deviation influencing the article representation and the user representation. In short, the connecting edges of the confounding factors M and N to the node I and the node U in fig. 1 are cut off, so as to cut off the influence of the confounding factors M and N on the exposure distribution of the article and the user. The backdoor adjusting technology [3] is used for converting an expression P (Y | do (X)) containing a do operator into a form which can be represented by data, and the specific conversion mode is as follows:

wherein Y represents the final result, equivalent to the click probability in the recommendation system; x represents a reason, and is equivalent to an item representation node or a user representation node; z represents a confounding factor, equivalent to item popularity or user liveness.

In the present invention, the representation P (Y | do (U, I)) by the back door adjustment technique is as follows:

where P (·) is a probability, P (m) is a probability of item popularity, and P (n) is a probability of user activity.

In the step 3, the BPR loss function [4] is a loss function commonly used in the recommendation system sequencing task, and the formula is as follows:

l _BPR ＝-ln(σ(y _ui -y _uj ))

where σ (-) denotes the sigmoid function, y _ui And y _uj Respectively, the scores of positive and negative examples, wherein the positive example i is an item interacted by the user u, and the negative example j is an item not interacted by the user u. The purpose of the BPR penalty function training is to learn how much a given user u prefers positive samples i over negative samples j. The information of the popularity of the goods and the liveness of the user is added in the training process, so that the existing BPR loss function can be improved, namely the improved BPR loss function can be trained by using the information of the popularity of the goods and the liveness of the user, and the formula is as follows:

l＝l ₁ +l ₂ +λ·||Θ|| ²

wherein l ₁ Representing the popularity of items training the preference of a given user u for different items; l. the ₂ Representing the preference degrees of different users for the given item i in the activity training of the users; u is a user representation, i is an article representation interacted with by the user u, j is an article representation not interacted with by the user u, and v represents a user representation not interacted with the article i.

Indicating the popularity of item i at time t,

representing the activity of the user u at the moment t; σ (-) denotes the sigmoid function. λ is L ₂ And the regular term coefficient is used for controlling the overfitting. By minimizing loss contentNumber l to determine the parameter Θ and thus the user embedding vector and the item embedding vector.

Representing a historical interaction data set.

In step 4, the adjusting of the popularity deviation refers to intervening in the inference stage to obtain a development trend more consistent with the popularity. The intervention popularity is to use a mathematical formula to predict the popularity of the test set so as to alleviate the problem that data are not independently and identically distributed between the training set and the test set, and finally obtain the ranking of recommended articles. Generally, there are two types of mathematical formulas. First, the prediction formula is as follows:

herein, the

Representing the popularity of the last stage in the training set. That is, the popularity of the last stage in the training set is utilized as a prediction of the popularity of the items in the test set.

Secondly, the prediction formula is as follows:

herein, the

Indicating the popularity of the last stage in the training set,

representing the popularity of the penultimate stage in the training set. Alpha is a super parameter used to control the popularity offset.

The invention relates to a recommendation method based on double-view angle correction, which more comprehensively analyzes the deviation in the observation data, utilizes a backdoor adjustment technology in causal inference to remove the deviation on a test set in a training stage according to the analyzed deviation, and adjusts the deviation according to the distribution of the test set in an inference stage, so that the final recommendation result can reflect the real preference of a user, and the personalized experience of the user and the potential income of an article provider are improved.

The invention also provides a recommendation system for realizing the recommendation method, and the system comprises: the cause and effect estimation method comprises a cause and effect diagram construction module, a cause and effect estimation calculation module and a cause and effect estimation adjustment module.

Specifically, the cause-and-effect graph construction module analyzes whether the popularity deviation and the activity deviation of the user exist in the data by using the popularity deviation and the activity deviation, and constructs a cause-and-effect graph conforming to a data generation mechanism;

the cause and effect estimator calculation module parameterizes the cause and effect estimator P (Y | do (U, I)) and then fits training data by using a recommendation model comprising a matrix decomposition model or a LightGCN model;

the cause and effect estimator adjusting module predicts the popularity of the articles on the test set and adjusts the trained cause and effect estimator, namely adjusts the prediction result.

The beneficial effects of the invention include: by removing the popularity deviation in the training stage and adjusting the popularity deviation in the testing stage, the problem that the distribution of off-line training data and on-line testing data in the recommendation system is inconsistent is solved. Meanwhile, the deviation existing in the data is considered from the aspects of articles and users, and the prediction performance of the recommendation system is improved. Compared with the prior art, the prediction performance of the recommendation system is respectively 5.02%,19.70% and 6.94% in the Recall @20 indexes of three large-scale real data sets of double _ movie, ml10m and amazon _ book, and is respectively 7.55%,14.36% and 10.21% in the NDCG @20 indexes.

In addition, when the influence of the head user on the final recommendation performance is researched, the method provided by the invention has stronger capturing capability on the objects interacted by the active user. According to the invention, the Influence Rate (IRTU) of head Users is considered on an amazon _ book data set, namely, the most active K items interacted with by the Users are taken out from a training set, the recommended times of the items in a test set are taken as indexes, and the first 20 items interacted with by the Users are selected for statistics. As shown in FIG. 4, the IRTU value of DBDA is higher than that of PDA at different K values.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to these drawings without creative efforts.

FIG. 1 is a causal graph of the present invention that considers item popularity and user liveness.

Fig. 2 is a schematic diagram of the training process of the present invention.

FIG. 3 is a schematic diagram of a prediction process according to the present invention.

FIG. 4 is a bar graph of the head user influence level of the present invention on the amazon _ book data set.

FIG. 5 is a flow chart of the operation of the present invention.

Detailed Description

The invention is described in further detail with reference to the following specific examples and the accompanying drawings. The procedures, conditions, experimental methods and the like for carrying out the present invention are common knowledge and common general knowledge in the art except for the contents specifically mentioned below, and the present invention is not particularly limited.

The invention provides a recommendation method based on double-view-angle deviation correction, which comprises the following steps of firstly, constructing a cause-and-effect graph conforming to a data generation mechanism; then, defining a causal estimator capable of accurately representing the real interest of the user by utilizing a backdoor adjustment technology in causal inference; then, learning the embedded vectors of the articles and the users by using the article popularity and the user activity in the training stage to learn the real user interests; finally, the popularity deviation is adjusted in the deduction stage, potential popular articles on the test set are found, the over-fitting problem of the existing recommendation system is solved, and accurate recommendation according with the current preference of the user is achieved.

Specifically, the method comprises the steps of firstly constructing a cause-and-effect graph according with a data generation mechanism through a priori knowledge, secondly defining a cause-and-effect estimator capable of reflecting real preference of a user based on the cause-and-effect graph in a training phase and estimating the cause-and-effect estimator by using data, and finally finding potentially popular articles on a test set by adjusting popularity in an inference phase and recommending articles which the user may like to the user.

The priori knowledge comprises the fact that article popularity deviation and user activity deviation exist in interaction data of the recommendation system.

Step 1: a causal graph as shown in figure 1 was constructed. The causal graph shows two mixed factors of the popularity of the goods and the activity of the user, so that the decoupling effect is achieved. Where M denotes item popularity and N denotes user liveness. The decoupling means that the popularity of the item and the activity of the user are independent of each other.

Step 2: two-view depolarization is performed during the training phase. As shown in the schematic training flow diagram of the present invention in fig. 2, wherein the right arrow represents the item-side training phase, the left arrow represents the user-side training phase,

notation represents the inner product of the tensor, weight sharing during training, e _i Is an embedded vector of the article, e _u In the form of an embedded vector for the user,

indicating the popularity of item i at time t,

representing the activity of user u at time t. The aim of the training stage is to remove the influence of the popularity of the item and the liveness of the user, so as to obtain the real interest of the user on the given item, namely, the final click probability is only the embodiment of the real interest of the user and is not subject to the real interest of the userTo other confounding factors. Here, the do operator is used, P (Y | do (U, I)) is used to replace the conventional P (Y | U, I) to represent the final click probability, and the back door adjustment technique [3] is used]P (Y | do (U, I)) is represented by data, i.e.

Where P (-) is the probability, P (m) is the probability of item popularity, and P (n) is the probability of user activity. Note that the second equation of the above equation holds true because of the causal graph constructed in the present invention, as shown in fig. 1. The popularity of the article and the liveness of the user are independent.

Step 2.1: p (Y | U, I, M, N) is estimated. P (Y | U, I, M, N) represents the popularity of a given (U, I) pair and item I

And liveness of user u

Probability of occurrence of the lower click behavior. Setting the parameter of the conditional probability as theta, and interacting the data set through the history

And BPR loss function [4]]To learn the unknown parameters:

wherein

Representing a historical data set, namely interaction between a user and an article obtained by history, j representing a negative sample of the article by the user u, and sigma (·) representing a sigmoid function.

Will be as follows

Parameterization and transformation into a problem that can be solved by the invention:

wherein ELU' is an Exponential Linear Unit (ELU) activation function [6]A variant of (4). f. of _Θ (u, i) represents a user-item matching model, the invention employs a classical Matrix Factorization (MF) model [5 ]]The data were fitted. To ensure a matching function f _Θ The value of (. Cndot.) is always positive:

wherein the content of the first and second substances,

and

all values of (A) are positive numbers; if gamma is ₁ And gamma ₂ Zero, indicating that item popularity and user liveness are not functional; if gamma is ₁ And gamma ₂ The greater the impact of item popularity and user liveness. While ELU' (. Cndot.) is a monotonic function, which ensures monotonicity of the probability function.

Step 2.2: the loss function is refined based on the dual view angle. According to the parameterized causal estimator, the popularity of the goods and the activity of the user are independent. The parameterized causal estimator is substituted into the traditional BPR loss function, so that the fact that the popularity of the articles and the activity of the user are decoupled is easy to find, and the activity of the user has the same value on both sides of the minus sign when the BPR loss function is used for parameter training, so that the training cannot be influenced. In other words, the loss function directly constructed in this way does not utilize the information of the user activity, so that the purpose of removing the user activity cannot be achieved. Therefore, to solve this problem, the present invention improves the conventional BPR loss function to replace the original parameter training method. The invention utilizes the information of the popularity of the article and the liveness of the user to respectively construct two loss functions:

l＝l ₁ +l ₂ +λ·||Θ|| ²

wherein l ₁ The preference degrees, l, of a given user u to different articles are trained by using the popularity of the articles ₂ The user activity is utilized to train the preference degree of different users for a given item i. L is the complete loss function and λ is L ₂ And the regular term coefficient is used for controlling the overfitting. The parameter Θ is determined by minimizing the loss function l.

Step 2.3: estimate P (Y | do (U, I)). The popularity m of the article and the activity n of the user need to be summed, but the space is huge and the traversal is difficult. As can be seen from step 2.1, the item popularity and the user activity have been decoupled, and then:

where E (-) represents the expectation, which is generally considered a constant. In the recommendation problem, the final task is a task of ordering the items, so the invention focuses more on the relative value of the click probability rather than the absolute value. Therefore, P (Y | do (U, I)) can be directly estimated by ELU' (U · I).

In summary, the invention uses

Training with parameters using modified version of BPR loss function, fitting historical interaction data

And the popularity of the articles and the activity of the users can be utilized in the training process, and finally, an ELU' (u & i) is used for carrying out a sequencing task. The invention names the method as Double Bias Deconfiguration (DBD), namely respectively deflective two confounding factors of article popularity and user activity.

And step 3: the popularity bias is adjusted during the inference phase. FIG. 3 is a schematic diagram of the prediction flow of the present invention, wherein the predicted value is reached _DBD The path(s) represents the prediction mode of the traditional recommendation method, and the predicted value is reached _DBDA The path of (a) represents the prediction mode of the proposed method of the present invention,

the notation denotes the inner product of the tensor. P (Y | do (U, I)) is used during the training phase to avoid the accumulation of popularity bias. However, in the inference phase, the present invention improves recommendation performance by exploiting the popularity bias to find potentially popular items on the test set. It is noted that the invention does not require intervention in user activity, since in the sorting task, the intervention in user activity does not change the relative value of the click probability, i.e. does not change the order of the recommendation list. The inference phase then intervenes only on the item popularity:

wherein the content of the first and second substances,

indicating the popularity of items on the test set. According to the back-door adjustment technique, the probability of intervention may be directly equal to the conditional probability, since there is no back-door path between Y and M and Y and N.

Step 3.1: and predicting the popularity. Here, the

What represents the popularity of items on the test set, the present invention uses mathematical formulas to predict it. The prediction formula is as follows:

a first kind,

Herein, the

Representing the popularity of the last stage in the training set. That is, the popularity of the last stage in the training set is utilized as a prediction of the popularity of the items in the test set. This formula corresponds to the prediction method (a) in the embodiment.

A second kind,

Herein, the

The popularity of the last phase in the training set is indicated. Alpha is a super parameter used to control the popularity offset. This formula corresponds to the prediction method (b) in the embodiment.

Step 3.2: p (Y | do (U, I), do (M), do (N)) is estimated. And in the inference stage, the popularity deviation is utilized to adjust the prediction result:

the method is also trained with the improved BPR loss function in step 2, but the invention adjusts the popularity of the goods in the inference phase. The method is named as Double Bias Demixing and Adjusting (DBDA), namely, two confounding factors of the article popularity and the user activity are respectively deflexed in a training stage, and the article popularity is adjusted in an inference stage.

Examples

Referring to the flowchart of the overall method of fig. 5, the training of the recommendation model is performed as follows:

step 1: data sets are collected and data is preprocessed. The double Movie is data collected from the national bean website and contains the user's rating of the Movie within 10 years. The present embodiment takes the data after 2010 and takes all scored interactions as positive samples. Items and users with an interaction volume less than 10 are filtered out. The resulting interaction volume was 7,174,218, the number of users 47,890, and the number of items 26,047. In this embodiment, the data is divided into 10 time phases, each of which has the same time interval, and the last phase randomly takes 30% of the interaction data of the users as a verification set and 70% of the interaction data of the users as a test set. Users and items that appeared only at the last stage were not included in the experiment.

And 2, step: and training the processed data. Sampling method before training: for a given (u, i) pair, an item j which is not interacted with by a user u and a user v which is not interacted with an item i are randomly sampled to serve as negative samples, and a quadruple (u, v, i, j) is formed. Iterative training is then performed.

And 3, step 3: based on the parameters trained in step 2, in the inference stage, the popularity is adjusted according to two popularity prediction formulas defined previously.

Experiments are performed on a double Movie dataset, and comparison is performed with BPRMF, PD and PDA models. The experimental results are shown in table 1, where Recall is Recall, precision is Precision, hit Ratio (HR) is Hit Ratio of item in the test set appearing in the top-K recommendation list, and NDCG is relative order between positive and negative samples in the top-K recommendation list:

TABLE 1

Table 1 presents a comparison of recommended performance for 5 different methods on the double Movie dataset. The first 3 methods remove the popularity bias during the training phase. The latter 2 methods are to adjust the ranking with the popularity bias in the inference stage, and two popularity prediction methods are provided, one is to use the popularity of the previous stage as the predicted value of the popularity in the testing stage, corresponding to method (a) in the table, and the other is to use the popularity prediction formula defined previously, corresponding to method (b) in the table.

For each evaluation index, the bold in the table indicates the best results, and the underlining indicates suboptimal results. Looking first at the first 3 methods for each data set, it can be seen that the method of the present invention is superior to the other methods in each index. Looking again at the latter 2 methods, the method of the present invention is also overall superior to the comparative method PDA. Particularly, under the evaluation index of NDCG @ K, the method disclosed by the invention has a far better effect than that of PDA. In the double Movie data set, relative increase values of prediction methods (a) and (b) were 5.02% and 6.17%, respectively, on the Recall @20 index by the DBDA method, relative to the PDA method; relative increases for prediction methods (a) and (b) were 7.55% and 10.00% on the NDCG @20 index, respectively. Because the main idea of the NDCG index is that items preferred by the user are ranked in front of the recommendation list to a greater extent than behind the recommendation list, the method of the present invention predicts items with a higher preference rank, which better meets the recommendation requirements.

In the similar method, the influence of each popularity prediction formula on the result is compared. It can be seen that: either PDA or DBDA methods of the present invention.

The sensitivity to the predictive popularity formula is greater. In practical applications, a popularity prediction formula suitable for the data may be determined by the validation set. For the proposed method of the present invention, for example: the prediction method (a) was more effective on the double Movie dataset.

Reference to the literature

[1]CHEN J W,WANG X,FENG F L,et al.Bias issues and solutions in recommender system[C]//Proceedings of the 15th ACM Conference on Recommender Systems,Amsterdam,Sep 27-Oct 1,2021.NewYork:ACM,2021:825-827.

[2]ZHANG Y,FENG F L,HE X N,et al.Causal intervention for leveraging popularity bias in recommendation[C]//Proceedings of the 44th ACM SIGIR Conference on Research and Development in Information Retrieval,Montreal,Jul 11-15,2021.New York:ACM,2021:11-20.

[3]PEARL J.Causality[M].2nd ed.New York:Cambridge university press,2009:65-106.

[4]RENDLE S,FREUDENTHALER C,GANTNER Z,et al.BPR:bayesian personalized ranking from implicit feedback[C]//Proceedings of the 25th Conference on Uncertainty in Artificial Intelligence,Montreal,Jun 18-21,2009.AUAI,2009:452-461.

[5]KOREN Y,BELL R,VOLINSKY C.Matrix factorization techniques for recommender systems[J]//the IEEE Computer Society,2009,42(8):30-37.

[6]CLEVERT DA,UNTERTHINER T,HOCHREITER S.Fast and accurate deep network learning by exponential linear units(elus)[C]//Proceedings of the 4th International Conference on Learning Representations,San Juan,May 2-4,2016.OpenReview.net,2016.

The protection of the present invention is not limited to the above embodiments. Variations and advantages that may occur to those skilled in the art may be incorporated into the invention without departing from the spirit and scope of the inventive concept, and the scope of the appended claims is intended to be protected.

Claims (10)

1. A recommendation method based on double-view-angle deviation correction is characterized in that the recommendation method considers two view angles of an article and a user at the same time, and comprises the following steps:

step 1: quantitatively analyzing whether the data has article popularity deviation and user activity deviation or not, and constructing a causal graph according with a data generation mechanism;

step 2: based on the causal graph constructed in the step 1, in a training phase, a do operator is used, P (Y | do (U, I)) is used for replacing traditional P (Y | U, I), the influence of the popularity deviation of the item characteristics and the user characteristics is eliminated, the real preference of the user to the item is expressed, and then a back door adjustment technology is used for estimating a P (Y | do (U, I)) causal estimator through data;

and step 3: parameterizing the causal estimator in the step 2, so that the causal estimator can be recovered by using existing data and trained, and improving a Bayesian personalized sorting (BPR) loss function, so that a recommendation model can be subjected to depolarization by using information of two aspects of article popularity and user activity; the recommendation model comprises a matrix decomposition model and a LightGCN model;

and 4, step 4: in the inference stage, items that are potentially popular on the test set are found by adjusting the popularity bias.

2. The recommendation method based on dual-view rectification according to claim 1, wherein the causal graph in step 1 is composed of a user representation node U, an article representation node I and a final click probability Y, wherein U and I are input as models, and Y is obtained through training, so that the nodes U and I point to the node Y at the same time; on the basis, nodes M and N are defined and respectively represent article popularity and user activity, M serves as a confounding factor and simultaneously points to nodes I and Y, and N is a confounding factor and simultaneously points to nodes U and Y in the same way.

3. The dual-view based recommendation method of claim 2, wherein the popularity M of the item refers to the popularity of the item, and the popularity of the item i at the time t is defined as

Wherein

Representing the interaction quantity of an item I at the time t, and representing the item set by an item representation node I;

the user activity N refers to the activity degree of the user, and the popularity of the user u at the moment t is defined as

Wherein

Representing the interaction quantity of the user U at the time t, wherein the user representation node U represents a set of users;

the user characterization node U represents an embedded vector of a user;

the item characterization node I represents an embedded vector of an item;

the final click probability Y represents the click probability of the user on the item after the two factors of the user and the item act.

4. The recommendation method based on dual-view rectification according to claim 1, characterized in that in step 1, quantitative analysis is performed on whether article popularity deviation exists in the data, the data set is divided according to a time sequence, the data set is divided into ten stages according to interaction time, each stage has the same time interval, so that article popularity in different time periods is measured, and on the basis, a trend of the article popularity along with time is depicted by using popularity offset, so that whether the article popularity deviation exists in the data is quantitatively analyzed;

the method comprises the steps of carrying out quantitative analysis on whether user activity deviation exists in data or not, dividing a data set according to a time sequence, dividing the data set into ten stages according to interaction time, wherein each stage has the same time interval, and describing the change trend of the user activity along with the time by using activity offset on the basis, so that whether the user activity deviation exists in the data or not is quantitatively analyzed.

5. The dual-perspective deskew-based recommendation method of claim 4, wherein said item popularity offset is expressed as Jensen-Shannon divergence metric

Wherein t is ₁ And t ₂ The two time periods are divided into two time periods,

is t ₁ The probability distribution of the items at a time,

is t ₂ Probability distribution of the items at the moment;

the user activity offset is expressed as Jensen-Shannon divergence measurement

Wherein t is ₁ And t ₂ The two time periods are divided into two time periods,

is t ₁ The probability distribution of the users at the moment in time,

is t ₂ Probability distribution of users at the moment.

6. The recommendation method based on dual-view rectification according to claim 1, wherein in step 2, the do operator cuts off the influence of confounding factors M and N on the exposure distribution of the articles and the users by cutting off the connecting edges of confounding factors M and N in the cause-and-effect graph to the nodes I and U, so as to remove the popularity deviation affecting the article characterization and the user characterization.

7. The method as claimed in claim 1, wherein in step 2, the back door adjustment technique converts the expression P (Y | do (X)) containing the do operator into a form represented by data, and the conversion is as follows:

P(Y＝y|do(X＝x))＝∑ _z P(Y＝y|X＝x，Z＝z)P(Z＝z)，

wherein Y represents the final result, which is equivalent to the click probability in the recommendation system; x represents a reason, and is equivalent to an article representation node or a user representation node; z represents a confounding factor, equivalent to item popularity or user liveness;

the representation of P (Y | do (U, I)) using the back door adjustment technique is as follows:

P(Y|do(U，I))＝∑ _m，n P(Y|U，I，m，n)P(m，n)＝∑ _m，n P(Y|U，I，m，n)P(m)P(n)，

where P (·) is a probability, P (m) is a probability of item popularity, and P (n) is a probability of user activity.

8. The recommendation method based on dual-view rectification as claimed in claim 1, wherein in step 3, the information of article popularity and user activity is utilized to perform the depolarization to improve the BPR loss function, and the formula of the improved BPR loss function is as follows:

l＝l ₁ +l ₂ +λ·||Θ|| ² ，

wherein l ₁ Representing the popularity of the item training the preference degree of a given user u for different items; l ₂ Representing the degree of preference of different users for the given item i for the activity training of the users; u is a user representation, i is an article representation interacted by the user u, j is an article representation not interacted by the user u, and v represents a user representation not interacted with the article i;

indicating the popularity of item i at time t,

representing the activity of the user u at the moment t; σ (-) denotes a sigmoid function; λ is L ₂ Regular term coefficients, controlling overfitting; determining a parameter Θ by minimizing a loss function l, thereby determining a user embedding vector and an item embedding vector;

representing a historical interaction data set.

9. The recommendation method based on dual-view rectification as claimed in claim 1, wherein the adjusting of the popularity deviation in step 4 refers to intervening in the inference stage to obtain a development trend more in line with the popularity, that is, using a mathematical formula to predict the popularity on the test set to alleviate the problem of non-independent and uniform distribution of data between the training set and the test set, and finally obtaining the ranking of the recommended articles;

the mathematical formula comprises

Or

Wherein, the first and the second end of the pipe are connected with each other,

representing the popularity of the last stage in the training set;

representing the popularity of the penultimate stage in the training set; alpha is a hyper parameter used to control the popularity offset.

10. A recommendation system for implementing the recommendation method according to any one of claims 1-9, characterized in that the system comprises: the cause and effect estimation method comprises a cause and effect graph building module, a cause and effect estimation quantity calculating module and a cause and effect estimation quantity adjusting module;

the cause-and-effect graph construction module analyzes whether the popularity deviation and the activity deviation of the user exist in the data or not by using the popularity deviation and the activity deviation, and constructs a cause-and-effect graph according with a data generation mechanism;

the cause and effect estimator calculation module parameterizes the cause and effect estimator P (Y | do (U, I)) and then fits training data by using a recommendation model comprising a matrix decomposition model or a LightGCN model;

the cause and effect estimator adjusting module predicts the popularity of the articles on the test set and adjusts the trained cause and effect estimator, namely adjusts the prediction result.

Images